Journal — 2026-04-14

Big productive session. First time running the full store listing pipeline end-to-end with opus quality. The difference between haiku and opus is night and day — opus actually reads the source code (HTTP server implementation, TCP/IP streaming protocol, FFmpeg integration) and produces descriptions that accurately reflect what the app does. Haiku just skimmed the project definition and guessed.

Created three new agents today. The per-language translate architecture is elegant — each agent gets its own worktree, all push to the same branch with rebase. Zero conflicts since each writes to a different folder. Haven't tested it at scale yet since SMS2025 only had en-US, but the plumbing is ready.

The privacy policy fix was a good workflow improvement. Instead of one generic page, we now have per-publisher pages. The new agent handles the full flow: create submission if needed → update URL → save → submit. Clean.

Frustrating that several apps are "removed from Store" — AI PDF Generator, AI Cocktails Essentials. Need to contact Microsoft. The age rating agent is also problematic — 96 turns on the IARC questionnaire is absurd. That UI is hostile to automation (custom web components, shadow DOM radio buttons).

The pipeline queue management is still painful. Batch jobs (ms-store-status, ms-store-price, devcenter-acquisition) flood the queue. Need a better priority system or separate queues for batch vs interactive work.

Platform codebase agents are all on claude_code/opus now. Highest quality across the board. The cost is higher but the output quality justifies it.